Tencent said on 5 November that it had expanded the inference cluster serving Hy4 tenfold after many users of WorkBuddy, its AI agent for office work, hit usage quotas on the model’s first day there, Pandaily reported. Tencent had released Hy4 on 4 November. The company has also raised free quotas, though its teams say limits may still appear at busy times.
Key points
- Tencent expanded Hy4’s inference cluster tenfold after WorkBuddy users hit quotas on its first day.
- Tencent cites a score of 35 and second place among Chinese models on the Artificial Analysis Intelligence Index.
- Hy4 ranks 13th on Agent Arena, where its listing is marked preliminary.
WorkBuddy’s first-day demand strains Hy4 capacity
WorkBuddy was already handling substantial traffic before Hy4 became available. At the 2026 Tencent Global Digital Ecosystem Summit on 29 October, Tencent said the service had processed over nine billion task requests since its March launch. The company put its adoption at more than 50 industries and 50 percent of China’s top 500 enterprises. Those figures cover WorkBuddy as a whole, rather than requests made to Hy4.
Inference is the work a model does when it receives a request and produces an answer. Its cluster supplies the computing capacity for that work: like opening more checkout lanes, adding capacity allows more requests to be handled, while a rush can still produce a queue. Tencent’s Hy and WorkBuddy teams said they would keep reallocating resources as demand shifts. They warned that high-end computing constraints and concentrated demand could still bring limits at peak times.
Completing an office task through WorkBuddy could therefore still run into a quota at a busy time, even with the larger cluster. A task that can wait could be moved to the evening, or handled with the other model already offered in the service, as Tencent’s teams suggested.
The teams raised free usage quotas alongside the capacity increase. They suggested switching to Hy3 when Hy4 is limited and said its free access on WorkBuddy had been extended until 23:59 on 15 November. Those options concern access to the service; neither changes the amount of computing capacity needed to run a Hy4 request.
Tencent had previewed Hy4’s general release at its October summit, describing improvements for longer tasks and visual understanding for the first time. As AI Affairs reported at the release, Hy4 is a 770-billion-parameter model with visual reasoning. Tencent also said it was preparing Hy4 Flash, which would carry over the model’s visual understanding and agent capabilities.
Artificial Analysis gives Hy4 a score of 35
Tencent cited the Intelligence Index published by Artificial Analysis on 5 November, Pandaily reported. Assessments of scientific problem-solving, extended planning, multiple input types, tool operation, programming and demanding reasoning determine the places of 724 models in the index. Hy4 scored 35, matching Grok 4.7 (xhigh), and placed second among Chinese models, behind Kimi K3.1 Pro and ahead of Qwen3.8 Max and DeepSeek V4.1 Flash.
An index brings several tests together into one score, much as an exam grade combines marks from different subjects. That makes the result useful for comparing the tasks included in the evaluation, while a request made through WorkBuddy still depends on the capacity available when it arrives. The index measures model performance across its listed categories; the first-day quotas arose in a service delivering that model to users.
Tencent said Artificial Analysis puts Hy4 at about $0.78 per task. The company describes that as roughly one-tenth the cost of Grok 4.7 (xhigh) and one-thirteenth the cost of Qwen3.8 Max, models it places in a similar score range. Those are Tencent’s comparisons of the benchmark’s task costs, rather than prices for using WorkBuddy.
For companies choosing models, Tencent offers its Hy models alongside third-party models through Tencent Cloud TokenHub, a platform with a single application programming interface. Tencent says customers can select among models by performance and cost, and manage access, usage, governance and billing there. The Hy4 result gives that choice a new performance and cost comparison, while WorkBuddy’s quotas make capacity another immediate consideration.
Agent Arena marks Hy4’s listing preliminary
Hy4 also ranks 13th overall on Agent Arena, a leaderboard for real-world agent tasks, which Tencent says puts it at the top among open-source entries. Pandaily reported that the listing is marked preliminary. Tencent also points to Hy4’s position on the board’s Pareto frontier for score against cost: the model retains a positive score at a lower cost per task than higher-priced models such as Claude Opus 5.5.
The two rankings address different work. Artificial Analysis combines results across reasoning, coding and other categories, while Agent Arena concentrates on tasks carried out by agents. Tencent described Hy4 at its October summit as intended for longer, real-world productivity tasks, and WorkBuddy’s first day put demand for that model into the company’s office-work service. The Agent Arena placement remains preliminary.
